Designing with AI-enabled Patient Archetypes from Online Health Communities

Where behavioral data hides

Chronic disease management increasingly demands care that extends beyond biomedical indicators to incorporate psychosocial and behavioral factors that shape day-to-day self-management. Standardized care processes often miss patient-specific goals, motivations, coping strategies, and practical constraints that influence engagement and adherence.

Clinicians know that understanding patients better leads to better care. They lack systematic ways to capture behavioral insights before consultations. The challenge is not just about data. It is about timing, scale, and the limits of clinical intuition.

I try to instinctively know, but I definitely am wrong in a few cases. It's instinct. I think I'm better now than I was 10 years ago just by experience. But I do remember cases where I thought, okay, this patient wants me to do this. And then after a couple of days realized, okay, no, I just hit the wrong button here. If I know the patient better, I can give them better treatment options.

— Clinician, Erasmus MC

The data was hiding in plain sight

To understand how to capture behavioral insights at scale, I explored where patient narratives actually exist. Online health communities contain large volumes of patient-authored narratives that describe symptoms, worries, uncertainty, side effects, emotional burden, and everyday challenges. The behavioral data we needed was not missing—it was hiding in plain sight.

These narratives offer rich dimensions of experience that are only partially captured in structured clinical documentation. They describe a patient's individual psychosocial and behavioral factors. Methodological work shows these accounts can yield credible and actionable insights when data sources, sampling, and analytical procedures are transparently documented and made traceable. Recent studies show that large language models can transform free form medical narratives into structured data, enabling scalable analysis while preserving clinically relevant detail.

264,358 patient posts analyzed · 3 chronic conditions · 6,500 curated narratives · 130 clinician validated topics

From 264,358 stories to 8 archetypes

I developed a reproducible, human-in-the-loop pipeline that extracts patient goals, motivations, and challenges from online narratives and organizes them into behavior-focused, traceable archetypes that clinicians can inspect, interrogate, and revise.

Rather than treating LLMs as back-end analytical systems that feed into opaque decision support, we configured the model as a co-analyst whose outputs remain traceable to source narratives and open to clinical interpretation and revision. The archetypes function as design materials, not black-box predictions.

The design process involved building a traceable pipeline from patient narratives to actionable archetypes, then designing a dashboard that integrates these insights into clinical workflows. Every step was validated with clinicians to ensure clinical relevance and usability.

AI proposes, clinician decides

The dashboard lives inside the electronic health record system. AI derived archetypes inform care modules presented as editable personalization blocks. The system follows an "AI proposes, clinician disposes" approach. Clinicians always have final control.

The dashboard layers behavioral insights alongside conventional clinical information. Clinicians can see not just what a patient has, but how they approach their care. Every insight remains traceable to source narratives. This preserves accountability and supports clinical interpretation.

Decisions that shaped trust

AI proposes, clinician disposes

The system never makes autonomous decisions. It suggests care modules based on behavioral patterns. Clinicians always have final say. As one expert noted, "It gives suggestions, not decisions. That's important. Just giving information or suggestions makes it a non MDR regulated, non medical device. So you stay outside that scope." This was not just about safety. It was about trust and preserving clinical agency. Every insight is traceable to source quotes. Clinicians can verify the reasoning and adapt suggestions to their clinical judgment.

Behavior over diagnosis

Traditional systems organize patients by disease. Our system layers behavioral insights on top of clinical data. Two lung cancer patients might need completely different communication styles and care approaches. The diagnosis stays the same. The care pathway adapts to how each patient thinks, copes, and engages with their care. As one pulmonologist observed, "This is really something new, a basis to know what patients do in daily life and adjust treatment or motivate them to walk more. It really helps clinicians."

Embedded, not adjacent

We designed the dashboard to live inside HIX, Erasmus MC's EHR system. It is not a separate tool. Clinicians should not have to context switch between systems. Behavioral insights appear right where patient data already lives. They integrate seamlessly into existing workflows. The Head of Oncology emphasized the value of embedding this in the electronic patient file.

Traceable and interrogable

Every behavioral insight links back to source narratives. Clinicians can click through to see the patient quotes that informed an archetype assignment. This enables them to verify, question, and sometimes resist AI derived behavioral framings. As one expert noted, "I would definitely look at behavioral data before consultation. That might help shape the conversation."

It's tremendous work actually. If this could be embedded in our electronic patient file, that would really be a big help.

— Head of Oncology, Erasmus MC

What the wards proved

This project proved that behavioral personalization at scale is possible. We took 264,358 scattered patient stories and turned them into a systematic framework that clinicians can actually use while preserving their agency. The framework supports richer patient-clinician conversations and enables more informed care decisions within the constraints of 15-minute consultations.

Expert walkthroughs revealed that behavioral insights help clinicians prepare more effectively for consultations, enabling them to shape conversations around patient-specific needs and motivations. The system supports shared decision-making by presenting behavioral context alongside clinical data, helping clinicians understand not just what a patient has, but how they approach their care. This positions AI as a partner rather than a replacement in chronic care planning, preserving clinical agency while augmenting decision-making capabilities.

I get tired looking at HIX all day. This is calmer. To keep the overview, this is very good. It helps to have a summary of what the care pathway looks like.

— Clinician, Erasmus MC

What I would do differently

This work contributes to interaction design and the HCI community by showing how AI-enabled design materials can be made accountable and revisable within care infrastructures, linking patient voices to clinical planning.

• Earlier and deeper patient involvement in validating archetypes.
• Longitudinal validation of clinical outcomes and experiences.
• Broader disease coverage to test generalizability.